Feature Selection
Selection procedures that do not leak the test set into the choice.
Markets are low signal-to-noise, non-stationary and adversarial — the conditions under which machine learning most easily fools its operator. We treat validation as the hard part of the work, not a formality after training.
On a feature selection engagement this means starting from what you already have — an idea, a spreadsheet, a running system, or a set of results that do not add up — and working from there rather than insisting on a rebuild. Where the honest answer is that the work is not worth doing, that is what you will be told.
What you get
- Feature engineering with strict point-in-time discipline
- Purged, embargoed cross-validation appropriate to financial series
- Honest baselines, so a model must beat something before it is deployed
- Drift monitoring and retraining policy defined before go-live
How the work runs
Scope
A short call and a written scope: what the system must do, what data it needs, what counts as done. Fixed price where the scope allows it.
Build
Work in reviewable increments, with running code you can see rather than a status update. Assumptions are surfaced as they arise, not at the end.
Test
Independent testing against the acceptance criteria agreed in the scope, including the failure cases.
Handover
Source code, documentation and a walkthrough. You own the result and can maintain it without us.
Tell us what you are trying to build
Describe the problem and we will tell you plainly whether we can help, roughly what it would take, and what it would cost. If it is not a fit, we will say so.
Discuss My ML Trading Project